Supported Formats

Use this page to choose an entry point from the files you already have. NepTrainKit converts common structure files and training outputs to its internal Structure representation for visualization, filtering, candidate generation, and export.

Structure files

File	Recommended entry point	Typical use
`.xyz` / `.extxyz`	`Make Dataset` or `NEP Dataset Display`	Starting structures, a candidate pool, or a training set
`POSCAR` / `CONTCAR`	`Make Dataset`	Generate a candidate pool from an initial crystal structure
`CIF`	`Make Dataset`	Build a candidate set from a crystal structure
ASE `.traj`	`NEP Dataset Display`	Inspect an existing trajectory or convert structures
DeepMD `deepmd/npy` directory	`NEP Dataset Display`	Inspect or export data in DeepMD format

Import the data into Make Dataset when you want to generate more structures. Use NEP Dataset Display when the next task is to inspect, remove, filter, or export a subset.

Importing a DeepMD directory requires type.raw, type_map.raw, set.* / box.npy, and set.* / coord.npy, with consistent frame counts, atom counts, element mappings, cells, and coordinates. Missing or inconsistent contracts fail explicitly instead of continuing with placeholder elements, an identity cell, or incomplete frames.

EXTXYZ data contract

NEP Dataset Display validates atom counts, Lattice, Properties, element symbols, per-atom column counts, and logical fields against the extended XYZ contract. A frame fails explicitly if its declared atom count differs from the number of rows, columns are missing, an element or PBC value is invalid, or a periodic cell is degenerate; missing data is never padded with zeros. Consistent with the GPUMD/NEP EXTXYZ convention, frames without pbc default to fully periodic T T T. Nonperiodic or partially periodic data must declare pbc explicitly in the header.

NEP training files

File	Purpose
`nep.txt`	NEP model used for prediction, retrospective inspection, or prescreening
`energy_*.out`	Energy predictions or reference labels
`force_*.out`	Force predictions or reference labels
`virial_*.out`	Virial predictions or reference labels
`stress_*.out`	Stress predictions or reference labels
`descriptor*.out`	Descriptor output used for visualization or sampling analysis

After training, load the training structures together with these outputs in NEP Dataset Display. You can inspect scatter plots, locate the largest-error structures, and export a subset for the next iteration.

VASP

OUTCAR: reads the cell, coordinates, forces, stress, and virial.
XDATCAR: reads the cell and coordinates for each frame.

VASP results are commonly used to convert DFT labels into training structures or to inspect label quality in NEP Dataset Display.

Advanced structure formats

The Advanced / experimental structure files filter under NEP Dataset Display → Open File… also exposes CP2K output and n2p2 input.data. Both importers identify the format from file contents; a matching extension alone does not make a damaged file valid.

CP2K single-point output

Main .out / .log files containing a CP2K signature are supported.
This entry point currently supports only three-dimensional periodic single-point output containing one coordinate block and one force block. Convert multi-step optimization or MD histories to EXTXYZ first; multiple coordinate blocks are never concatenated into one structure.
The importer reads CELL| Vector a/b/c [angstrom], MODULE QUICKSTEP: ATOMIC COORDINATES IN ANGSTROM, Hartree energy, forces in atomic units, and stress in GPa, then converts them to eV, Å, and eV/Å³.
Import fails explicitly when the complete cell or coordinates are missing, the force count does not match, the cell is degenerate, or multiple coordinate or force blocks are present.
The current CP2K main-output entry point records pbc="T T T"; do not use it for PERIODIC NONE or partially periodic calculations. Export those calculations from CP2K as EXTXYZ with explicit Lattice and pbc, then import that file.

The CP2K geometry-optimization documentation likewise recommends trajectory or final-structure files for multi-step geometries; the main output is not a complete trajectory container.

n2p2 `input.data`

Files named input.data and files with .data or .cfg extensions are supported.
Every structure block must be enclosed by begin and end. Each atom row must contain position, element, charge, per-atom energy, and a three-dimensional force.
n2p2 CFG values are read in Bohr and Hartree, then converted to Å, eV, and eV/Å.
Either zero or three lattice rows form a valid contract. One or two lattice rows, damaged numeric values, invalid elements, or a missing end marker cause the entire import to fail.

LAMMPS

dump
lammpstrj

Orthogonal and triclinic cells and the common coordinate columns are supported. After import, spot-check the cell, element types, and coordinate units. Every frame must provide complete BOX BOUNDS boundary flags, one complete coordinate triplet (xs/ys/zs, x/y/z, or xu/yu/zu), and an element or type column. A dump containing only numeric type values requires a complete type-number-to-element mapping from the import UI or caller; NepTrainKit never substitutes pseudo-elements for unknown types. Truncated atom rows, invalid elements, missing mappings, or degenerate cells fail the entire import instead of returning a partial trajectory.

Recommended dump format with spins

When using spin LAMMPS, use the following column names to output magnetic moments:

compute spin all property/atom sp spx spy spz fmx fmy fmz fx fy fz
dump dpgen_dump all custom 100 traj.dump id type x y z c_spin[1] c_spin[2] c_spin[3] c_spin[4] c_spin[5] c_spin[6] c_spin[7]

The importer treats c_spin[1] as the moment magnitude and c_spin[2], c_spin[3], and c_spin[4] as the three direction components, then reconstructs the per-atom spin vector:

\[ \mathbf s_i=c_{\mathrm{spin}[1]} \left(c_{\mathrm{spin}[2]},c_{\mathrm{spin}[3]},c_{\mathrm{spin}[4]}\right). \]

c_spin[5], c_spin[6], and c_spin[7] are magnetic-force components and are not written to spin:R:3. If any column in c_spin[1:4] is missing, the importer does not create a spin field for that frame.

Export

NEP Dataset Display can export the current data in two main forms:

xyz / extxyz: suitable for continuing in GPUMD, ASE, or a Make Dataset workflow.
deepmd/npy: suitable for DeepMD-style data management. Settings → Keep DeepMD subfolders controls whether the imported subdirectory hierarchy is preserved.

When exporting deepmd/npy, NepTrainKit preserves fully periodic or fully nonperiodic PBC; a fully nonperiodic directory receives a nopbc marker. Structures sharing a Config_type must have consistent elements and atom count, PBC, per-atom fields, and energy/virial coverage. Because the standard DeepMD directory format cannot unambiguously represent partial PBC or mixed contracts in one directory, such inputs fail before any files are written. Per-atom integer, logical, and string fields round-trip with their original types, while atom order is canonicalized by type_map.

Large datasets and import behavior

For a large dataset, the interface may sample or defer rendering for responsiveness. Selection and export still operate on the complete dataset.
You can cancel an import while it is running.
Use descriptive names for files from different stages, such as candidate_pool.xyz, candidate_pool_clean.xyz, and dft_labeled.xyz.